fs/nfs/filelayout/filelayout.c

   1 /*
   2  *  Module for the pnfs nfs4 file layout driver.
   3  *  Defines all I/O and Policy interface operations, plus code
   4  *  to register itself with the pNFS client.
   5  *
   6  *  Copyright (c) 2002
   7  *  The Regents of the University of Michigan
   8  *  All Rights Reserved
   9  *
  10  *  Dean Hildebrand <dhildebz@umich.edu>
  11  *
  12  *  Permission is granted to use, copy, create derivative works, and
  13  *  redistribute this software and such derivative works for any purpose,
  14  *  so long as the name of the University of Michigan is not used in
  15  *  any advertising or publicity pertaining to the use or distribution
  16  *  of this software without specific, written prior authorization. If
  17  *  the above copyright notice or any other identification of the
  18  *  University of Michigan is included in any copy of any portion of
  19  *  this software, then the disclaimer below must also be included.
  20  *
  21  *  This software is provided as is, without representation or warranty
  22  *  of any kind either express or implied, including without limitation
  23  *  the implied warranties of merchantability, fitness for a particular
  24  *  purpose, or noninfringement.  The Regents of the University of
  25  *  Michigan shall not be liable for any damages, including special,
  26  *  indirect, incidental, or consequential damages, with respect to any
  27  *  claim arising out of or in connection with the use of the software,
  28  *  even if it has been or is hereafter advised of the possibility of
  29  *  such damages.
  30  */
  31
  32 #include <linux/nfs_fs.h>
  33 #include <linux/nfs_page.h>
  34 #include <linux/module.h>
  35
  36 #include <linux/sunrpc/metrics.h>
  37
  38 #include "../nfs4session.h"
  39 #include "../internal.h"
  40 #include "../delegation.h"
  41 #include "filelayout.h"
  42 #include "../nfs4trace.h"
  43
  44 #define NFSDBG_FACILITY         NFSDBG_PNFS_LD
  45
  46 MODULE_LICENSE("GPL");
  47 MODULE_AUTHOR("Dean Hildebrand <dhildebz@umich.edu>");
  48 MODULE_DESCRIPTION("The NFSv4 file layout driver");
  49
  50 #define FILELAYOUT_POLL_RETRY_MAX     (15*HZ)
  51
  52 static loff_t
  53 filelayout_get_dense_offset(struct nfs4_filelayout_segment *flseg,
  54                             loff_t offset)
  55 {
  56         u32 stripe_width = flseg->stripe_unit * flseg->dsaddr->stripe_count;
  57         u64 stripe_no;
  58         u32 rem;
  59
  60         offset -= flseg->pattern_offset;
  61         stripe_no = div_u64(offset, stripe_width);
  62         div_u64_rem(offset, flseg->stripe_unit, &rem);
  63
  64         return stripe_no * flseg->stripe_unit + rem;
  65 }
  66
  67 /* This function is used by the layout driver to calculate the
  68  * offset of the file on the dserver based on whether the
  69  * layout type is STRIPE_DENSE or STRIPE_SPARSE
  70  */
  71 static loff_t
  72 filelayout_get_dserver_offset(struct pnfs_layout_segment *lseg, loff_t offset)
  73 {
  74         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
  75
  76         switch (flseg->stripe_type) {
  77         case STRIPE_SPARSE:
  78                 return offset;
  79
  80         case STRIPE_DENSE:
  81                 return filelayout_get_dense_offset(flseg, offset);
  82         }
  83
  84         BUG();
  85 }
  86
  87 static void filelayout_reset_write(struct nfs_pgio_header *hdr)
  88 {
  89         struct rpc_task *task = &hdr->task;
  90
  91         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
  92                 dprintk("%s Reset task %5u for i/o through MDS "
  93                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
  94                         hdr->task.tk_pid,
  95                         hdr->inode->i_sb->s_id,
  96                         (unsigned long long)NFS_FILEID(hdr->inode),
  97                         hdr->args.count,
  98                         (unsigned long long)hdr->args.offset);
  99
 100                 task->tk_status = pnfs_write_done_resend_to_mds(hdr);
 101         }
 102 }
 103
 104 static void filelayout_reset_read(struct nfs_pgio_header *hdr)
 105 {
 106         struct rpc_task *task = &hdr->task;
 107
 108         if (!test_and_set_bit(NFS_IOHDR_REDO, &hdr->flags)) {
 109                 dprintk("%s Reset task %5u for i/o through MDS "
 110                         "(req %s/%llu, %u bytes @ offset %llu)\n", __func__,
 111                         hdr->task.tk_pid,
 112                         hdr->inode->i_sb->s_id,
 113                         (unsigned long long)NFS_FILEID(hdr->inode),
 114                         hdr->args.count,
 115                         (unsigned long long)hdr->args.offset);
 116
 117                 task->tk_status = pnfs_read_done_resend_to_mds(hdr);
 118         }
 119 }
 120
 121 static void filelayout_fenceme(struct inode *inode, struct pnfs_layout_hdr *lo)
 122 {
 123         if (!test_and_clear_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
 124                 return;
 125         pnfs_return_layout(inode);
 126 }
 127
 128 static int filelayout_async_handle_error(struct rpc_task *task,
 129                                          struct nfs4_state *state,
 130                                          struct nfs_client *clp,
 131                                          struct pnfs_layout_segment *lseg)
 132 {
 133         struct pnfs_layout_hdr *lo = lseg->pls_layout;
 134         struct inode *inode = lo->plh_inode;
 135         struct nfs_server *mds_server = NFS_SERVER(inode);
 136         struct nfs4_deviceid_node *devid = FILELAYOUT_DEVID_NODE(lseg);
 137         struct nfs_client *mds_client = mds_server->nfs_client;
 138         struct nfs4_slot_table *tbl = &clp->cl_session->fc_slot_table;
 139
 140         if (task->tk_status >= 0)
 141                 return 0;
 142
 143         switch (task->tk_status) {
 144         /* MDS state errors */
 145         case -NFS4ERR_DELEG_REVOKED:
 146         case -NFS4ERR_ADMIN_REVOKED:
 147         case -NFS4ERR_BAD_STATEID:
 148                 if (state == NULL)
 149                         break;
 150                 nfs_remove_bad_delegation(state->inode);
 151         case -NFS4ERR_OPENMODE:
 152                 if (state == NULL)
 153                         break;
 154                 if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 155                         goto out_bad_stateid;
 156                 goto wait_on_recovery;
 157         case -NFS4ERR_EXPIRED:
 158                 if (state != NULL) {
 159                         if (nfs4_schedule_stateid_recovery(mds_server, state) < 0)
 160                                 goto out_bad_stateid;
 161                 }
 162                 nfs4_schedule_lease_recovery(mds_client);
 163                 goto wait_on_recovery;
 164         /* DS session errors */
 165         case -NFS4ERR_BADSESSION:
 166         case -NFS4ERR_BADSLOT:
 167         case -NFS4ERR_BAD_HIGH_SLOT:
 168         case -NFS4ERR_DEADSESSION:
 169         case -NFS4ERR_CONN_NOT_BOUND_TO_SESSION:
 170         case -NFS4ERR_SEQ_FALSE_RETRY:
 171         case -NFS4ERR_SEQ_MISORDERED:
 172                 dprintk("%s ERROR %d, Reset session. Exchangeid "
 173                         "flags 0x%x\n", __func__, task->tk_status,
 174                         clp->cl_exchange_flags);
 175                 nfs4_schedule_session_recovery(clp->cl_session, task->tk_status);
 176                 break;
 177         case -NFS4ERR_DELAY:
 178         case -NFS4ERR_GRACE:
 179                 rpc_delay(task, FILELAYOUT_POLL_RETRY_MAX);
 180                 break;
 181         case -NFS4ERR_RETRY_UNCACHED_REP:
 182                 break;
 183         /* Invalidate Layout errors */
 184         case -NFS4ERR_PNFS_NO_LAYOUT:
 185         case -ESTALE:           /* mapped NFS4ERR_STALE */
 186         case -EBADHANDLE:       /* mapped NFS4ERR_BADHANDLE */
 187         case -EISDIR:           /* mapped NFS4ERR_ISDIR */
 188         case -NFS4ERR_FHEXPIRED:
 189         case -NFS4ERR_WRONG_TYPE:
 190                 dprintk("%s Invalid layout error %d\n", __func__,
 191                         task->tk_status);
 192                 /*
 193                  * Destroy layout so new i/o will get a new layout.
 194                  * Layout will not be destroyed until all current lseg
 195                  * references are put. Mark layout as invalid to resend failed
 196                  * i/o and all i/o waiting on the slot table to the MDS until
 197                  * layout is destroyed and a new valid layout is obtained.
 198                  */
 199                 pnfs_destroy_layout(NFS_I(inode));
 200                 rpc_wake_up(&tbl->slot_tbl_waitq);
 201                 goto reset;
 202         /* RPC connection errors */
 203         case -ECONNREFUSED:
 204         case -EHOSTDOWN:
 205         case -EHOSTUNREACH:
 206         case -ENETUNREACH:
 207         case -EIO:
 208         case -ETIMEDOUT:
 209         case -EPIPE:
 210                 dprintk("%s DS connection error %d\n", __func__,
 211                         task->tk_status);
 212                 nfs4_mark_deviceid_unavailable(devid);
 213                 set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags);
 214                 rpc_wake_up(&tbl->slot_tbl_waitq);
 215                 /* fall through */
 216         default:
 217 reset:
 218                 dprintk("%s Retry through MDS. Error %d\n", __func__,
 219                         task->tk_status);
 220                 return -NFS4ERR_RESET_TO_MDS;
 221         }
 222 out:
 223         task->tk_status = 0;
 224         return -EAGAIN;
 225 out_bad_stateid:
 226         task->tk_status = -EIO;
 227         return 0;
 228 wait_on_recovery:
 229         rpc_sleep_on(&mds_client->cl_rpcwaitq, task, NULL);
 230         if (test_bit(NFS4CLNT_MANAGER_RUNNING, &mds_client->cl_state) == 0)
 231                 rpc_wake_up_queued_task(&mds_client->cl_rpcwaitq, task);
 232         goto out;
 233 }
 234
 235 /* NFS_PROTO call done callback routines */
 236
 237 static int filelayout_read_done_cb(struct rpc_task *task,
 238                                 struct nfs_pgio_header *hdr)
 239 {
 240         int err;
 241
 242         trace_nfs4_pnfs_read(hdr, task->tk_status);
 243         err = filelayout_async_handle_error(task, hdr->args.context->state,
 244                                             hdr->ds_clp, hdr->lseg);
 245
 246         switch (err) {
 247         case -NFS4ERR_RESET_TO_MDS:
 248                 filelayout_reset_read(hdr);
 249                 return task->tk_status;
 250         case -EAGAIN:
 251                 rpc_restart_call_prepare(task);
 252                 return -EAGAIN;
 253         }
 254
 255         return 0;
 256 }
 257
 258 /*
 259  * We reference the rpc_cred of the first WRITE that triggers the need for
 260  * a LAYOUTCOMMIT, and use it to send the layoutcommit compound.
 261  * rfc5661 is not clear about which credential should be used.
 262  */
 263 static void
 264 filelayout_set_layoutcommit(struct nfs_pgio_header *hdr)
 265 {
 266
 267         if (FILELAYOUT_LSEG(hdr->lseg)->commit_through_mds ||
 268             hdr->res.verf->committed != NFS_DATA_SYNC)
 269                 return;
 270
 271         pnfs_set_layoutcommit(hdr);
 272         dprintk("%s inode %lu pls_end_pos %lu\n", __func__, hdr->inode->i_ino,
 273                 (unsigned long) NFS_I(hdr->inode)->layout->plh_lwb);
 274 }
 275
 276 bool
 277 filelayout_test_devid_unavailable(struct nfs4_deviceid_node *node)
 278 {
 279         return filelayout_test_devid_invalid(node) ||
 280                 nfs4_test_deviceid_unavailable(node);
 281 }
 282
 283 static bool
 284 filelayout_reset_to_mds(struct pnfs_layout_segment *lseg)
 285 {
 286         struct nfs4_deviceid_node *node = FILELAYOUT_DEVID_NODE(lseg);
 287
 288         return filelayout_test_devid_unavailable(node);
 289 }
 290
 291 /*
 292  * Call ops for the async read/write cases
 293  * In the case of dense layouts, the offset needs to be reset to its
 294  * original value.
 295  */
 296 static void filelayout_read_prepare(struct rpc_task *task, void *data)
 297 {
 298         struct nfs_pgio_header *hdr = data;
 299
 300         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
 301                 rpc_exit(task, -EIO);
 302                 return;
 303         }
 304         if (filelayout_reset_to_mds(hdr->lseg)) {
 305                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 306                 filelayout_reset_read(hdr);
 307                 rpc_exit(task, 0);
 308                 return;
 309         }
 310         hdr->pgio_done_cb = filelayout_read_done_cb;
 311
 312         if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
 313                         &hdr->args.seq_args,
 314                         &hdr->res.seq_res,
 315                         task))
 316                 return;
 317         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
 318                         hdr->args.lock_context, FMODE_READ) == -EIO)
 319                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 320 }
 321
 322 static void filelayout_read_call_done(struct rpc_task *task, void *data)
 323 {
 324         struct nfs_pgio_header *hdr = data;
 325
 326         dprintk("--> %s task->tk_status %d\n", __func__, task->tk_status);
 327
 328         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
 329             task->tk_status == 0) {
 330                 nfs41_sequence_done(task, &hdr->res.seq_res);
 331                 return;
 332         }
 333
 334         /* Note this may cause RPC to be resent */
 335         hdr->mds_ops->rpc_call_done(task, data);
 336 }
 337
 338 static void filelayout_read_count_stats(struct rpc_task *task, void *data)
 339 {
 340         struct nfs_pgio_header *hdr = data;
 341
 342         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
 343 }
 344
 345 static void filelayout_read_release(void *data)
 346 {
 347         struct nfs_pgio_header *hdr = data;
 348         struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
 349
 350         filelayout_fenceme(lo->plh_inode, lo);
 351         nfs_put_client(hdr->ds_clp);
 352         hdr->mds_ops->rpc_release(data);
 353 }
 354
 355 static int filelayout_write_done_cb(struct rpc_task *task,
 356                                 struct nfs_pgio_header *hdr)
 357 {
 358         int err;
 359
 360         trace_nfs4_pnfs_write(hdr, task->tk_status);
 361         err = filelayout_async_handle_error(task, hdr->args.context->state,
 362                                             hdr->ds_clp, hdr->lseg);
 363
 364         switch (err) {
 365         case -NFS4ERR_RESET_TO_MDS:
 366                 filelayout_reset_write(hdr);
 367                 return task->tk_status;
 368         case -EAGAIN:
 369                 rpc_restart_call_prepare(task);
 370                 return -EAGAIN;
 371         }
 372
 373         filelayout_set_layoutcommit(hdr);
 374         return 0;
 375 }
 376
 377 /* Fake up some data that will cause nfs_commit_release to retry the writes. */
 378 static void prepare_to_resend_writes(struct nfs_commit_data *data)
 379 {
 380         struct nfs_page *first = nfs_list_entry(data->pages.next);
 381
 382         data->task.tk_status = 0;
 383         memcpy(&data->verf.verifier, &first->wb_verf,
 384                sizeof(data->verf.verifier));
 385         data->verf.verifier.data[0]++; /* ensure verifier mismatch */
 386 }
 387
 388 static int filelayout_commit_done_cb(struct rpc_task *task,
 389                                      struct nfs_commit_data *data)
 390 {
 391         int err;
 392
 393         trace_nfs4_pnfs_commit_ds(data, task->tk_status);
 394         err = filelayout_async_handle_error(task, NULL, data->ds_clp,
 395                                             data->lseg);
 396
 397         switch (err) {
 398         case -NFS4ERR_RESET_TO_MDS:
 399                 prepare_to_resend_writes(data);
 400                 return -EAGAIN;
 401         case -EAGAIN:
 402                 rpc_restart_call_prepare(task);
 403                 return -EAGAIN;
 404         }
 405
 406         if (data->verf.committed == NFS_UNSTABLE)
 407                 pnfs_commit_set_layoutcommit(data);
 408
 409         return 0;
 410 }
 411
 412 static void filelayout_write_prepare(struct rpc_task *task, void *data)
 413 {
 414         struct nfs_pgio_header *hdr = data;
 415
 416         if (unlikely(test_bit(NFS_CONTEXT_BAD, &hdr->args.context->flags))) {
 417                 rpc_exit(task, -EIO);
 418                 return;
 419         }
 420         if (filelayout_reset_to_mds(hdr->lseg)) {
 421                 dprintk("%s task %u reset io to MDS\n", __func__, task->tk_pid);
 422                 filelayout_reset_write(hdr);
 423                 rpc_exit(task, 0);
 424                 return;
 425         }
 426         if (nfs41_setup_sequence(hdr->ds_clp->cl_session,
 427                         &hdr->args.seq_args,
 428                         &hdr->res.seq_res,
 429                         task))
 430                 return;
 431         if (nfs4_set_rw_stateid(&hdr->args.stateid, hdr->args.context,
 432                         hdr->args.lock_context, FMODE_WRITE) == -EIO)
 433                 rpc_exit(task, -EIO); /* lost lock, terminate I/O */
 434 }
 435
 436 static void filelayout_write_call_done(struct rpc_task *task, void *data)
 437 {
 438         struct nfs_pgio_header *hdr = data;
 439
 440         if (test_bit(NFS_IOHDR_REDO, &hdr->flags) &&
 441             task->tk_status == 0) {
 442                 nfs41_sequence_done(task, &hdr->res.seq_res);
 443                 return;
 444         }
 445
 446         /* Note this may cause RPC to be resent */
 447         hdr->mds_ops->rpc_call_done(task, data);
 448 }
 449
 450 static void filelayout_write_count_stats(struct rpc_task *task, void *data)
 451 {
 452         struct nfs_pgio_header *hdr = data;
 453
 454         rpc_count_iostats(task, NFS_SERVER(hdr->inode)->client->cl_metrics);
 455 }
 456
 457 static void filelayout_write_release(void *data)
 458 {
 459         struct nfs_pgio_header *hdr = data;
 460         struct pnfs_layout_hdr *lo = hdr->lseg->pls_layout;
 461
 462         filelayout_fenceme(lo->plh_inode, lo);
 463         nfs_put_client(hdr->ds_clp);
 464         hdr->mds_ops->rpc_release(data);
 465 }
 466
 467 static void filelayout_commit_prepare(struct rpc_task *task, void *data)
 468 {
 469         struct nfs_commit_data *wdata = data;
 470
 471         nfs41_setup_sequence(wdata->ds_clp->cl_session,
 472                         &wdata->args.seq_args,
 473                         &wdata->res.seq_res,
 474                         task);
 475 }
 476
 477 static void filelayout_write_commit_done(struct rpc_task *task, void *data)
 478 {
 479         struct nfs_commit_data *wdata = data;
 480
 481         /* Note this may cause RPC to be resent */
 482         wdata->mds_ops->rpc_call_done(task, data);
 483 }
 484
 485 static void filelayout_commit_count_stats(struct rpc_task *task, void *data)
 486 {
 487         struct nfs_commit_data *cdata = data;
 488
 489         rpc_count_iostats(task, NFS_SERVER(cdata->inode)->client->cl_metrics);
 490 }
 491
 492 static void filelayout_commit_release(void *calldata)
 493 {
 494         struct nfs_commit_data *data = calldata;
 495
 496         data->completion_ops->completion(data);
 497         pnfs_put_lseg(data->lseg);
 498         nfs_put_client(data->ds_clp);
 499         nfs_commitdata_release(data);
 500 }
 501
 502 static const struct rpc_call_ops filelayout_read_call_ops = {
 503         .rpc_call_prepare = filelayout_read_prepare,
 504         .rpc_call_done = filelayout_read_call_done,
 505         .rpc_count_stats = filelayout_read_count_stats,
 506         .rpc_release = filelayout_read_release,
 507 };
 508
 509 static const struct rpc_call_ops filelayout_write_call_ops = {
 510         .rpc_call_prepare = filelayout_write_prepare,
 511         .rpc_call_done = filelayout_write_call_done,
 512         .rpc_count_stats = filelayout_write_count_stats,
 513         .rpc_release = filelayout_write_release,
 514 };
 515
 516 static const struct rpc_call_ops filelayout_commit_call_ops = {
 517         .rpc_call_prepare = filelayout_commit_prepare,
 518         .rpc_call_done = filelayout_write_commit_done,
 519         .rpc_count_stats = filelayout_commit_count_stats,
 520         .rpc_release = filelayout_commit_release,
 521 };
 522
 523 static enum pnfs_try_status
 524 filelayout_read_pagelist(struct nfs_pgio_header *hdr)
 525 {
 526         struct pnfs_layout_segment *lseg = hdr->lseg;
 527         struct nfs4_pnfs_ds *ds;
 528         struct rpc_clnt *ds_clnt;
 529         loff_t offset = hdr->args.offset;
 530         u32 j, idx;
 531         struct nfs_fh *fh;
 532
 533         dprintk("--> %s ino %lu pgbase %u req %Zu@%llu\n",
 534                 __func__, hdr->inode->i_ino,
 535                 hdr->args.pgbase, (size_t)hdr->args.count, offset);
 536
 537         /* Retrieve the correct rpc_client for the byte range */
 538         j = nfs4_fl_calc_j_index(lseg, offset);
 539         idx = nfs4_fl_calc_ds_index(lseg, j);
 540         ds = nfs4_fl_prepare_ds(lseg, idx);
 541         if (!ds)
 542                 return PNFS_NOT_ATTEMPTED;
 543
 544         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 545         if (IS_ERR(ds_clnt))
 546                 return PNFS_NOT_ATTEMPTED;
 547
 548         dprintk("%s USE DS: %s cl_count %d\n", __func__,
 549                 ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 550
 551         /* No multipath support. Use first DS */
 552         atomic_inc(&ds->ds_clp->cl_count);
 553         hdr->ds_clp = ds->ds_clp;
 554         hdr->ds_idx = idx;
 555         fh = nfs4_fl_select_ds_fh(lseg, j);
 556         if (fh)
 557                 hdr->args.fh = fh;
 558
 559         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
 560         hdr->mds_offset = offset;
 561
 562         /* Perform an asynchronous read to ds */
 563         nfs_initiate_pgio(ds_clnt, hdr,
 564                             &filelayout_read_call_ops, 0, RPC_TASK_SOFTCONN);
 565         return PNFS_ATTEMPTED;
 566 }
 567
 568 /* Perform async writes. */
 569 static enum pnfs_try_status
 570 filelayout_write_pagelist(struct nfs_pgio_header *hdr, int sync)
 571 {
 572         struct pnfs_layout_segment *lseg = hdr->lseg;
 573         struct nfs4_pnfs_ds *ds;
 574         struct rpc_clnt *ds_clnt;
 575         loff_t offset = hdr->args.offset;
 576         u32 j, idx;
 577         struct nfs_fh *fh;
 578
 579         /* Retrieve the correct rpc_client for the byte range */
 580         j = nfs4_fl_calc_j_index(lseg, offset);
 581         idx = nfs4_fl_calc_ds_index(lseg, j);
 582         ds = nfs4_fl_prepare_ds(lseg, idx);
 583         if (!ds)
 584                 return PNFS_NOT_ATTEMPTED;
 585
 586         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, hdr->inode);
 587         if (IS_ERR(ds_clnt))
 588                 return PNFS_NOT_ATTEMPTED;
 589
 590         dprintk("%s ino %lu sync %d req %Zu@%llu DS: %s cl_count %d\n",
 591                 __func__, hdr->inode->i_ino, sync, (size_t) hdr->args.count,
 592                 offset, ds->ds_remotestr, atomic_read(&ds->ds_clp->cl_count));
 593
 594         hdr->pgio_done_cb = filelayout_write_done_cb;
 595         atomic_inc(&ds->ds_clp->cl_count);
 596         hdr->ds_clp = ds->ds_clp;
 597         hdr->ds_idx = idx;
 598         fh = nfs4_fl_select_ds_fh(lseg, j);
 599         if (fh)
 600                 hdr->args.fh = fh;
 601         hdr->args.offset = filelayout_get_dserver_offset(lseg, offset);
 602
 603         /* Perform an asynchronous write */
 604         nfs_initiate_pgio(ds_clnt, hdr,
 605                                     &filelayout_write_call_ops, sync,
 606                                     RPC_TASK_SOFTCONN);
 607         return PNFS_ATTEMPTED;
 608 }
 609
 610 /*
 611  * filelayout_check_layout()
 612  *
 613  * Make sure layout segment parameters are sane WRT the device.
 614  * At this point no generic layer initialization of the lseg has occurred,
 615  * and nothing has been added to the layout_hdr cache.
 616  *
 617  */
 618 static int
 619 filelayout_check_layout(struct pnfs_layout_hdr *lo,
 620                         struct nfs4_filelayout_segment *fl,
 621                         struct nfs4_layoutget_res *lgr,
 622                         struct nfs4_deviceid *id,
 623                         gfp_t gfp_flags)
 624 {
 625         struct nfs4_deviceid_node *d;
 626         struct nfs4_file_layout_dsaddr *dsaddr;
 627         int status = -EINVAL;
 628
 629         dprintk("--> %s\n", __func__);
 630
 631         /* FIXME: remove this check when layout segment support is added */
 632         if (lgr->range.offset != 0 ||
 633             lgr->range.length != NFS4_MAX_UINT64) {
 634                 dprintk("%s Only whole file layouts supported. Use MDS i/o\n",
 635                         __func__);
 636                 goto out;
 637         }
 638
 639         if (fl->pattern_offset > lgr->range.offset) {
 640                 dprintk("%s pattern_offset %lld too large\n",
 641                                 __func__, fl->pattern_offset);
 642                 goto out;
 643         }
 644
 645         if (!fl->stripe_unit) {
 646                 dprintk("%s Invalid stripe unit (%u)\n",
 647                         __func__, fl->stripe_unit);
 648                 goto out;
 649         }
 650
 651         /* find and reference the deviceid */
 652         d = nfs4_find_get_deviceid(NFS_SERVER(lo->plh_inode), id,
 653                         lo->plh_lc_cred, gfp_flags);
 654         if (d == NULL)
 655                 goto out;
 656
 657         dsaddr = container_of(d, struct nfs4_file_layout_dsaddr, id_node);
 658         /* Found deviceid is unavailable */
 659         if (filelayout_test_devid_unavailable(&dsaddr->id_node))
 660                 goto out_put;
 661
 662         fl->dsaddr = dsaddr;
 663
 664         if (fl->first_stripe_index >= dsaddr->stripe_count) {
 665                 dprintk("%s Bad first_stripe_index %u\n",
 666                                 __func__, fl->first_stripe_index);
 667                 goto out_put;
 668         }
 669
 670         if ((fl->stripe_type == STRIPE_SPARSE &&
 671             fl->num_fh > 1 && fl->num_fh != dsaddr->ds_num) ||
 672             (fl->stripe_type == STRIPE_DENSE &&
 673             fl->num_fh != dsaddr->stripe_count)) {
 674                 dprintk("%s num_fh %u not valid for given packing\n",
 675                         __func__, fl->num_fh);
 676                 goto out_put;
 677         }
 678
 679         status = 0;
 680 out:
 681         dprintk("--> %s returns %d\n", __func__, status);
 682         return status;
 683 out_put:
 684         nfs4_fl_put_deviceid(dsaddr);
 685         goto out;
 686 }
 687
 688 static void filelayout_free_fh_array(struct nfs4_filelayout_segment *fl)
 689 {
 690         int i;
 691
 692         for (i = 0; i < fl->num_fh; i++) {
 693                 if (!fl->fh_array[i])
 694                         break;
 695                 kfree(fl->fh_array[i]);
 696         }
 697         kfree(fl->fh_array);
 698         fl->fh_array = NULL;
 699 }
 700
 701 static void
 702 _filelayout_free_lseg(struct nfs4_filelayout_segment *fl)
 703 {
 704         filelayout_free_fh_array(fl);
 705         kfree(fl);
 706 }
 707
 708 static int
 709 filelayout_decode_layout(struct pnfs_layout_hdr *flo,
 710                          struct nfs4_filelayout_segment *fl,
 711                          struct nfs4_layoutget_res *lgr,
 712                          struct nfs4_deviceid *id,
 713                          gfp_t gfp_flags)
 714 {
 715         struct xdr_stream stream;
 716         struct xdr_buf buf;
 717         struct page *scratch;
 718         __be32 *p;
 719         uint32_t nfl_util;
 720         int i;
 721
 722         dprintk("%s: set_layout_map Begin\n", __func__);
 723
 724         scratch = alloc_page(gfp_flags);
 725         if (!scratch)
 726                 return -ENOMEM;
 727
 728         xdr_init_decode_pages(&stream, &buf, lgr->layoutp->pages, lgr->layoutp->len);
 729         xdr_set_scratch_buffer(&stream, page_address(scratch), PAGE_SIZE);
 730
 731         /* 20 = ufl_util (4), first_stripe_index (4), pattern_offset (8),
 732          * num_fh (4) */
 733         p = xdr_inline_decode(&stream, NFS4_DEVICEID4_SIZE + 20);
 734         if (unlikely(!p))
 735                 goto out_err;
 736
 737         memcpy(id, p, sizeof(*id));
 738         p += XDR_QUADLEN(NFS4_DEVICEID4_SIZE);
 739         nfs4_print_deviceid(id);
 740
 741         nfl_util = be32_to_cpup(p++);
 742         if (nfl_util & NFL4_UFLG_COMMIT_THRU_MDS)
 743                 fl->commit_through_mds = 1;
 744         if (nfl_util & NFL4_UFLG_DENSE)
 745                 fl->stripe_type = STRIPE_DENSE;
 746         else
 747                 fl->stripe_type = STRIPE_SPARSE;
 748         fl->stripe_unit = nfl_util & ~NFL4_UFLG_MASK;
 749
 750         fl->first_stripe_index = be32_to_cpup(p++);
 751         p = xdr_decode_hyper(p, &fl->pattern_offset);
 752         fl->num_fh = be32_to_cpup(p++);
 753
 754         dprintk("%s: nfl_util 0x%X num_fh %u fsi %u po %llu\n",
 755                 __func__, nfl_util, fl->num_fh, fl->first_stripe_index,
 756                 fl->pattern_offset);
 757
 758         /* Note that a zero value for num_fh is legal for STRIPE_SPARSE.
 759          * Futher checking is done in filelayout_check_layout */
 760         if (fl->num_fh >
 761             max(NFS4_PNFS_MAX_STRIPE_CNT, NFS4_PNFS_MAX_MULTI_CNT))
 762                 goto out_err;
 763
 764         if (fl->num_fh > 0) {
 765                 fl->fh_array = kcalloc(fl->num_fh, sizeof(fl->fh_array[0]),
 766                                        gfp_flags);
 767                 if (!fl->fh_array)
 768                         goto out_err;
 769         }
 770
 771         for (i = 0; i < fl->num_fh; i++) {
 772                 /* Do we want to use a mempool here? */
 773                 fl->fh_array[i] = kmalloc(sizeof(struct nfs_fh), gfp_flags);
 774                 if (!fl->fh_array[i])
 775                         goto out_err_free;
 776
 777                 p = xdr_inline_decode(&stream, 4);
 778                 if (unlikely(!p))
 779                         goto out_err_free;
 780                 fl->fh_array[i]->size = be32_to_cpup(p++);
 781                 if (sizeof(struct nfs_fh) < fl->fh_array[i]->size) {
 782                         printk(KERN_ERR "NFS: Too big fh %d received %d\n",
 783                                i, fl->fh_array[i]->size);
 784                         goto out_err_free;
 785                 }
 786
 787                 p = xdr_inline_decode(&stream, fl->fh_array[i]->size);
 788                 if (unlikely(!p))
 789                         goto out_err_free;
 790                 memcpy(fl->fh_array[i]->data, p, fl->fh_array[i]->size);
 791                 dprintk("DEBUG: %s: fh len %d\n", __func__,
 792                         fl->fh_array[i]->size);
 793         }
 794
 795         __free_page(scratch);
 796         return 0;
 797
 798 out_err_free:
 799         filelayout_free_fh_array(fl);
 800 out_err:
 801         __free_page(scratch);
 802         return -EIO;
 803 }
 804
 805 static void
 806 filelayout_free_lseg(struct pnfs_layout_segment *lseg)
 807 {
 808         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 809
 810         dprintk("--> %s\n", __func__);
 811         nfs4_fl_put_deviceid(fl->dsaddr);
 812         /* This assumes a single RW lseg */
 813         if (lseg->pls_range.iomode == IOMODE_RW) {
 814                 struct nfs4_filelayout *flo;
 815
 816                 flo = FILELAYOUT_FROM_HDR(lseg->pls_layout);
 817                 flo->commit_info.nbuckets = 0;
 818                 kfree(flo->commit_info.buckets);
 819                 flo->commit_info.buckets = NULL;
 820         }
 821         _filelayout_free_lseg(fl);
 822 }
 823
 824 static int
 825 filelayout_alloc_commit_info(struct pnfs_layout_segment *lseg,
 826                              struct nfs_commit_info *cinfo,
 827                              gfp_t gfp_flags)
 828 {
 829         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 830         struct pnfs_commit_bucket *buckets;
 831         int size, i;
 832
 833         if (fl->commit_through_mds)
 834                 return 0;
 835
 836         size = (fl->stripe_type == STRIPE_SPARSE) ?
 837                 fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 838
 839         if (cinfo->ds->nbuckets >= size) {
 840                 /* This assumes there is only one IOMODE_RW lseg.  What
 841                  * we really want to do is have a layout_hdr level
 842                  * dictionary of <multipath_list4, fh> keys, each
 843                  * associated with a struct list_head, populated by calls
 844                  * to filelayout_write_pagelist().
 845                  * */
 846                 return 0;
 847         }
 848
 849         buckets = kcalloc(size, sizeof(struct pnfs_commit_bucket),
 850                           gfp_flags);
 851         if (!buckets)
 852                 return -ENOMEM;
 853         for (i = 0; i < size; i++) {
 854                 INIT_LIST_HEAD(&buckets[i].written);
 855                 INIT_LIST_HEAD(&buckets[i].committing);
 856                 /* mark direct verifier as unset */
 857                 buckets[i].direct_verf.committed = NFS_INVALID_STABLE_HOW;
 858         }
 859
 860         spin_lock(cinfo->lock);
 861         if (cinfo->ds->nbuckets >= size)
 862                 goto out;
 863         for (i = 0; i < cinfo->ds->nbuckets; i++) {
 864                 list_splice(&cinfo->ds->buckets[i].written,
 865                             &buckets[i].written);
 866                 list_splice(&cinfo->ds->buckets[i].committing,
 867                             &buckets[i].committing);
 868                 buckets[i].direct_verf.committed =
 869                         cinfo->ds->buckets[i].direct_verf.committed;
 870                 buckets[i].wlseg = cinfo->ds->buckets[i].wlseg;
 871                 buckets[i].clseg = cinfo->ds->buckets[i].clseg;
 872         }
 873         swap(cinfo->ds->buckets, buckets);
 874         cinfo->ds->nbuckets = size;
 875 out:
 876         spin_unlock(cinfo->lock);
 877         kfree(buckets);
 878         return 0;
 879 }
 880
 881 static struct pnfs_layout_segment *
 882 filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 883                       struct nfs4_layoutget_res *lgr,
 884                       gfp_t gfp_flags)
 885 {
 886         struct nfs4_filelayout_segment *fl;
 887         int rc;
 888         struct nfs4_deviceid id;
 889
 890         dprintk("--> %s\n", __func__);
 891         fl = kzalloc(sizeof(*fl), gfp_flags);
 892         if (!fl)
 893                 return NULL;
 894
 895         rc = filelayout_decode_layout(layoutid, fl, lgr, &id, gfp_flags);
 896         if (rc != 0 || filelayout_check_layout(layoutid, fl, lgr, &id, gfp_flags)) {
 897                 _filelayout_free_lseg(fl);
 898                 return NULL;
 899         }
 900         return &fl->generic_hdr;
 901 }
 902
 903 /*
 904  * filelayout_pg_test(). Called by nfs_can_coalesce_requests()
 905  *
 906  * Return 0 if @req cannot be coalesced into @pgio, otherwise return the number
 907  * of bytes (maximum @req->wb_bytes) that can be coalesced.
 908  */
 909 static size_t
 910 filelayout_pg_test(struct nfs_pageio_descriptor *pgio, struct nfs_page *prev,
 911                    struct nfs_page *req)
 912 {
 913         unsigned int size;
 914         u64 p_stripe, r_stripe;
 915         u32 stripe_offset;
 916         u64 segment_offset = pgio->pg_lseg->pls_range.offset;
 917         u32 stripe_unit = FILELAYOUT_LSEG(pgio->pg_lseg)->stripe_unit;
 918
 919         /* calls nfs_generic_pg_test */
 920         size = pnfs_generic_pg_test(pgio, prev, req);
 921         if (!size)
 922                 return 0;
 923
 924         /* see if req and prev are in the same stripe */
 925         if (prev) {
 926                 p_stripe = (u64)req_offset(prev) - segment_offset;
 927                 r_stripe = (u64)req_offset(req) - segment_offset;
 928                 do_div(p_stripe, stripe_unit);
 929                 do_div(r_stripe, stripe_unit);
 930
 931                 if (p_stripe != r_stripe)
 932                         return 0;
 933         }
 934
 935         /* calculate remaining bytes in the current stripe */
 936         div_u64_rem((u64)req_offset(req) - segment_offset,
 937                         stripe_unit,
 938                         &stripe_offset);
 939         WARN_ON_ONCE(stripe_offset > stripe_unit);
 940         if (stripe_offset >= stripe_unit)
 941                 return 0;
 942         return min(stripe_unit - (unsigned int)stripe_offset, size);
 943 }
 944
 945 static void
 946 filelayout_pg_init_read(struct nfs_pageio_descriptor *pgio,
 947                         struct nfs_page *req)
 948 {
 949         if (!pgio->pg_lseg)
 950                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 951                                            req->wb_context,
 952                                            0,
 953                                            NFS4_MAX_UINT64,
 954                                            IOMODE_READ,
 955                                            GFP_KERNEL);
 956         /* If no lseg, fall back to read through mds */
 957         if (pgio->pg_lseg == NULL)
 958                 nfs_pageio_reset_read_mds(pgio);
 959 }
 960
 961 static void
 962 filelayout_pg_init_write(struct nfs_pageio_descriptor *pgio,
 963                          struct nfs_page *req)
 964 {
 965         struct nfs_commit_info cinfo;
 966         int status;
 967
 968         if (!pgio->pg_lseg)
 969                 pgio->pg_lseg = pnfs_update_layout(pgio->pg_inode,
 970                                            req->wb_context,
 971                                            0,
 972                                            NFS4_MAX_UINT64,
 973                                            IOMODE_RW,
 974                                            GFP_NOFS);
 975         /* If no lseg, fall back to write through mds */
 976         if (pgio->pg_lseg == NULL)
 977                 goto out_mds;
 978         nfs_init_cinfo(&cinfo, pgio->pg_inode, pgio->pg_dreq);
 979         status = filelayout_alloc_commit_info(pgio->pg_lseg, &cinfo, GFP_NOFS);
 980         if (status < 0) {
 981                 pnfs_put_lseg(pgio->pg_lseg);
 982                 pgio->pg_lseg = NULL;
 983                 goto out_mds;
 984         }
 985         return;
 986 out_mds:
 987         nfs_pageio_reset_write_mds(pgio);
 988 }
 989
 990 static const struct nfs_pageio_ops filelayout_pg_read_ops = {
 991         .pg_init = filelayout_pg_init_read,
 992         .pg_test = filelayout_pg_test,
 993         .pg_doio = pnfs_generic_pg_readpages,
 994 };
 995
 996 static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 997         .pg_init = filelayout_pg_init_write,
 998         .pg_test = filelayout_pg_test,
 999         .pg_doio = pnfs_generic_pg_writepages,
1000 };
1001
1002 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
1003 {
1004         if (fl->stripe_type == STRIPE_SPARSE)
1005                 return nfs4_fl_calc_ds_index(&fl->generic_hdr, j);
1006         else
1007                 return j;
1008 }
1009
1010 /* The generic layer is about to remove the req from the commit list.
1011  * If this will make the bucket empty, it will need to put the lseg reference.
1012  * Note this is must be called holding the inode (/cinfo) lock
1013  */
1014 static void
1015 filelayout_clear_request_commit(struct nfs_page *req,
1016                                 struct nfs_commit_info *cinfo)
1017 {
1018         struct pnfs_layout_segment *freeme = NULL;
1019
1020         if (!test_and_clear_bit(PG_COMMIT_TO_DS, &req->wb_flags))
1021                 goto out;
1022         cinfo->ds->nwritten--;
1023         if (list_is_singular(&req->wb_list)) {
1024                 struct pnfs_commit_bucket *bucket;
1025
1026                 bucket = list_first_entry(&req->wb_list,
1027                                           struct pnfs_commit_bucket,
1028                                           written);
1029                 freeme = bucket->wlseg;
1030                 bucket->wlseg = NULL;
1031         }
1032 out:
1033         nfs_request_remove_commit_list(req, cinfo);
1034         pnfs_put_lseg_locked(freeme);
1035 }
1036
1037 static void
1038 filelayout_mark_request_commit(struct nfs_page *req,
1039                                struct pnfs_layout_segment *lseg,
1040                                struct nfs_commit_info *cinfo)
1041
1042 {
1043         struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
1044         u32 i, j;
1045         struct list_head *list;
1046         struct pnfs_commit_bucket *buckets;
1047
1048         if (fl->commit_through_mds) {
1049                 list = &cinfo->mds->list;
1050                 spin_lock(cinfo->lock);
1051                 goto mds_commit;
1052         }
1053
1054         /* Note that we are calling nfs4_fl_calc_j_index on each page
1055          * that ends up being committed to a data server.  An attractive
1056          * alternative is to add a field to nfs_write_data and nfs_page
1057          * to store the value calculated in filelayout_write_pagelist
1058          * and just use that here.
1059          */
1060         j = nfs4_fl_calc_j_index(lseg, req_offset(req));
1061         i = select_bucket_index(fl, j);
1062         spin_lock(cinfo->lock);
1063         buckets = cinfo->ds->buckets;
1064         list = &buckets[i].written;
1065         if (list_empty(list)) {
1066                 /* Non-empty buckets hold a reference on the lseg.  That ref
1067                  * is normally transferred to the COMMIT call and released
1068                  * there.  It could also be released if the last req is pulled
1069                  * off due to a rewrite, in which case it will be done in
1070                  * filelayout_clear_request_commit
1071                  */
1072                 buckets[i].wlseg = pnfs_get_lseg(lseg);
1073         }
1074         set_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1075         cinfo->ds->nwritten++;
1076
1077 mds_commit:
1078         /* nfs_request_add_commit_list(). We need to add req to list without
1079          * dropping cinfo lock.
1080          */
1081         set_bit(PG_CLEAN, &(req)->wb_flags);
1082         nfs_list_add_request(req, list);
1083         cinfo->mds->ncommit++;
1084         spin_unlock(cinfo->lock);
1085         if (!cinfo->dreq) {
1086                 inc_zone_page_state(req->wb_page, NR_UNSTABLE_NFS);
1087                 inc_bdi_stat(page_file_mapping(req->wb_page)->backing_dev_info,
1088                              BDI_RECLAIMABLE);
1089                 __mark_inode_dirty(req->wb_context->dentry->d_inode,
1090                                    I_DIRTY_DATASYNC);
1091         }
1092 }
1093
1094 static u32 calc_ds_index_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1095 {
1096         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1097
1098         if (flseg->stripe_type == STRIPE_SPARSE)
1099                 return i;
1100         else
1101                 return nfs4_fl_calc_ds_index(lseg, i);
1102 }
1103
1104 static struct nfs_fh *
1105 select_ds_fh_from_commit(struct pnfs_layout_segment *lseg, u32 i)
1106 {
1107         struct nfs4_filelayout_segment *flseg = FILELAYOUT_LSEG(lseg);
1108
1109         if (flseg->stripe_type == STRIPE_SPARSE) {
1110                 if (flseg->num_fh == 1)
1111                         i = 0;
1112                 else if (flseg->num_fh == 0)
1113                         /* Use the MDS OPEN fh set in nfs_read_rpcsetup */
1114                         return NULL;
1115         }
1116         return flseg->fh_array[i];
1117 }
1118
1119 static int filelayout_initiate_commit(struct nfs_commit_data *data, int how)
1120 {
1121         struct pnfs_layout_segment *lseg = data->lseg;
1122         struct nfs4_pnfs_ds *ds;
1123         struct rpc_clnt *ds_clnt;
1124         u32 idx;
1125         struct nfs_fh *fh;
1126
1127         idx = calc_ds_index_from_commit(lseg, data->ds_commit_index);
1128         ds = nfs4_fl_prepare_ds(lseg, idx);
1129         if (!ds)
1130                 goto out_err;
1131
1132         ds_clnt = nfs4_find_or_create_ds_client(ds->ds_clp, data->inode);
1133         if (IS_ERR(ds_clnt))
1134                 goto out_err;
1135
1136         dprintk("%s ino %lu, how %d cl_count %d\n", __func__,
1137                 data->inode->i_ino, how, atomic_read(&ds->ds_clp->cl_count));
1138         data->commit_done_cb = filelayout_commit_done_cb;
1139         atomic_inc(&ds->ds_clp->cl_count);
1140         data->ds_clp = ds->ds_clp;
1141         fh = select_ds_fh_from_commit(lseg, data->ds_commit_index);
1142         if (fh)
1143                 data->args.fh = fh;
1144         return nfs_initiate_commit(ds_clnt, data,
1145                                    &filelayout_commit_call_ops, how,
1146                                    RPC_TASK_SOFTCONN);
1147 out_err:
1148         prepare_to_resend_writes(data);
1149         filelayout_commit_release(data);
1150         return -EAGAIN;
1151 }
1152
1153 static int
1154 transfer_commit_list(struct list_head *src, struct list_head *dst,
1155                      struct nfs_commit_info *cinfo, int max)
1156 {
1157         struct nfs_page *req, *tmp;
1158         int ret = 0;
1159
1160         list_for_each_entry_safe(req, tmp, src, wb_list) {
1161                 if (!nfs_lock_request(req))
1162                         continue;
1163                 kref_get(&req->wb_kref);
1164                 if (cond_resched_lock(cinfo->lock))
1165                         list_safe_reset_next(req, tmp, wb_list);
1166                 nfs_request_remove_commit_list(req, cinfo);
1167                 clear_bit(PG_COMMIT_TO_DS, &req->wb_flags);
1168                 nfs_list_add_request(req, dst);
1169                 ret++;
1170                 if ((ret == max) && !cinfo->dreq)
1171                         break;
1172         }
1173         return ret;
1174 }
1175
1176 /* Note called with cinfo->lock held. */
1177 static int
1178 filelayout_scan_ds_commit_list(struct pnfs_commit_bucket *bucket,
1179                                struct nfs_commit_info *cinfo,
1180                                int max)
1181 {
1182         struct list_head *src = &bucket->written;
1183         struct list_head *dst = &bucket->committing;
1184         int ret;
1185
1186         ret = transfer_commit_list(src, dst, cinfo, max);
1187         if (ret) {
1188                 cinfo->ds->nwritten -= ret;
1189                 cinfo->ds->ncommitting += ret;
1190                 bucket->clseg = bucket->wlseg;
1191                 if (list_empty(src))
1192                         bucket->wlseg = NULL;
1193                 else
1194                         pnfs_get_lseg(bucket->clseg);
1195         }
1196         return ret;
1197 }
1198
1199 /* Move reqs from written to committing lists, returning count of number moved.
1200  * Note called with cinfo->lock held.
1201  */
1202 static int filelayout_scan_commit_lists(struct nfs_commit_info *cinfo,
1203                                         int max)
1204 {
1205         int i, rv = 0, cnt;
1206
1207         for (i = 0; i < cinfo->ds->nbuckets && max != 0; i++) {
1208                 cnt = filelayout_scan_ds_commit_list(&cinfo->ds->buckets[i],
1209                                                      cinfo, max);
1210                 max -= cnt;
1211                 rv += cnt;
1212         }
1213         return rv;
1214 }
1215
1216 /* Pull everything off the committing lists and dump into @dst */
1217 static void filelayout_recover_commit_reqs(struct list_head *dst,
1218                                            struct nfs_commit_info *cinfo)
1219 {
1220         struct pnfs_commit_bucket *b;
1221         struct pnfs_layout_segment *freeme;
1222         int i;
1223
1224 restart:
1225         spin_lock(cinfo->lock);
1226         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1227                 if (transfer_commit_list(&b->written, dst, cinfo, 0)) {
1228                         freeme = b->wlseg;
1229                         b->wlseg = NULL;
1230                         spin_unlock(cinfo->lock);
1231                         pnfs_put_lseg(freeme);
1232                         goto restart;
1233                 }
1234         }
1235         cinfo->ds->nwritten = 0;
1236         spin_unlock(cinfo->lock);
1237 }
1238
1239 /* filelayout_search_commit_reqs - Search lists in @cinfo for the head reqest
1240  *                                 for @page
1241  * @cinfo - commit info for current inode
1242  * @page - page to search for matching head request
1243  *
1244  * Returns a the head request if one is found, otherwise returns NULL.
1245  */
1246 static struct nfs_page *
1247 filelayout_search_commit_reqs(struct nfs_commit_info *cinfo, struct page *page)
1248 {
1249         struct nfs_page *freq, *t;
1250         struct pnfs_commit_bucket *b;
1251         int i;
1252
1253         /* Linearly search the commit lists for each bucket until a matching
1254          * request is found */
1255         for (i = 0, b = cinfo->ds->buckets; i < cinfo->ds->nbuckets; i++, b++) {
1256                 list_for_each_entry_safe(freq, t, &b->written, wb_list) {
1257                         if (freq->wb_page == page)
1258                                 return freq->wb_head;
1259                 }
1260                 list_for_each_entry_safe(freq, t, &b->committing, wb_list) {
1261                         if (freq->wb_page == page)
1262                                 return freq->wb_head;
1263                 }
1264         }
1265
1266         return NULL;
1267 }
1268
1269 static void filelayout_retry_commit(struct nfs_commit_info *cinfo, int idx)
1270 {
1271         struct pnfs_ds_commit_info *fl_cinfo = cinfo->ds;
1272         struct pnfs_commit_bucket *bucket;
1273         struct pnfs_layout_segment *freeme;
1274         int i;
1275
1276         for (i = idx; i < fl_cinfo->nbuckets; i++) {
1277                 bucket = &fl_cinfo->buckets[i];
1278                 if (list_empty(&bucket->committing))
1279                         continue;
1280                 nfs_retry_commit(&bucket->committing, bucket->clseg, cinfo);
1281                 spin_lock(cinfo->lock);
1282                 freeme = bucket->clseg;
1283                 bucket->clseg = NULL;
1284                 spin_unlock(cinfo->lock);
1285                 pnfs_put_lseg(freeme);
1286         }
1287 }
1288
1289 static unsigned int
1290 alloc_ds_commits(struct nfs_commit_info *cinfo, struct list_head *list)
1291 {
1292         struct pnfs_ds_commit_info *fl_cinfo;
1293         struct pnfs_commit_bucket *bucket;
1294         struct nfs_commit_data *data;
1295         int i;
1296         unsigned int nreq = 0;
1297
1298         fl_cinfo = cinfo->ds;
1299         bucket = fl_cinfo->buckets;
1300         for (i = 0; i < fl_cinfo->nbuckets; i++, bucket++) {
1301                 if (list_empty(&bucket->committing))
1302                         continue;
1303                 data = nfs_commitdata_alloc();
1304                 if (!data)
1305                         break;
1306                 data->ds_commit_index = i;
1307                 spin_lock(cinfo->lock);
1308                 data->lseg = bucket->clseg;
1309                 bucket->clseg = NULL;
1310                 spin_unlock(cinfo->lock);
1311                 list_add(&data->pages, list);
1312                 nreq++;
1313         }
1314
1315         /* Clean up on error */
1316         filelayout_retry_commit(cinfo, i);
1317         /* Caller will clean up entries put on list */
1318         return nreq;
1319 }
1320
1321 /* This follows nfs_commit_list pretty closely */
1322 static int
1323 filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
1324                            int how, struct nfs_commit_info *cinfo)
1325 {
1326         struct nfs_commit_data *data, *tmp;
1327         LIST_HEAD(list);
1328         unsigned int nreq = 0;
1329
1330         if (!list_empty(mds_pages)) {
1331                 data = nfs_commitdata_alloc();
1332                 if (data != NULL) {
1333                         data->lseg = NULL;
1334                         list_add(&data->pages, &list);
1335                         nreq++;
1336                 } else {
1337                         nfs_retry_commit(mds_pages, NULL, cinfo);
1338                         filelayout_retry_commit(cinfo, 0);
1339                         cinfo->completion_ops->error_cleanup(NFS_I(inode));
1340                         return -ENOMEM;
1341                 }
1342         }
1343
1344         nreq += alloc_ds_commits(cinfo, &list);
1345
1346         if (nreq == 0) {
1347                 cinfo->completion_ops->error_cleanup(NFS_I(inode));
1348                 goto out;
1349         }
1350
1351         atomic_add(nreq, &cinfo->mds->rpcs_out);
1352
1353         list_for_each_entry_safe(data, tmp, &list, pages) {
1354                 list_del_init(&data->pages);
1355                 if (!data->lseg) {
1356                         nfs_init_commit(data, mds_pages, NULL, cinfo);
1357                         nfs_initiate_commit(NFS_CLIENT(inode), data,
1358                                             data->mds_ops, how, 0);
1359                 } else {
1360                         struct pnfs_commit_bucket *buckets;
1361
1362                         buckets = cinfo->ds->buckets;
1363                         nfs_init_commit(data, &buckets[data->ds_commit_index].committing, data->lseg, cinfo);
1364                         filelayout_initiate_commit(data, how);
1365                 }
1366         }
1367 out:
1368         cinfo->ds->ncommitting = 0;
1369         return PNFS_ATTEMPTED;
1370 }
1371 static struct nfs4_deviceid_node *
1372 filelayout_alloc_deviceid_node(struct nfs_server *server,
1373                 struct pnfs_device *pdev, gfp_t gfp_flags)
1374 {
1375         struct nfs4_file_layout_dsaddr *dsaddr;
1376
1377         dsaddr = nfs4_fl_alloc_deviceid_node(server, pdev, gfp_flags);
1378         if (!dsaddr)
1379                 return NULL;
1380         return &dsaddr->id_node;
1381 }
1382
1383 static void
1384 filelayout_free_deveiceid_node(struct nfs4_deviceid_node *d)
1385 {
1386         nfs4_fl_free_deviceid(container_of(d, struct nfs4_file_layout_dsaddr, id_node));
1387 }
1388
1389 static struct pnfs_layout_hdr *
1390 filelayout_alloc_layout_hdr(struct inode *inode, gfp_t gfp_flags)
1391 {
1392         struct nfs4_filelayout *flo;
1393
1394         flo = kzalloc(sizeof(*flo), gfp_flags);
1395         return flo != NULL ? &flo->generic_hdr : NULL;
1396 }
1397
1398 static void
1399 filelayout_free_layout_hdr(struct pnfs_layout_hdr *lo)
1400 {
1401         kfree(FILELAYOUT_FROM_HDR(lo));
1402 }
1403
1404 static struct pnfs_ds_commit_info *
1405 filelayout_get_ds_info(struct inode *inode)
1406 {
1407         struct pnfs_layout_hdr *layout = NFS_I(inode)->layout;
1408
1409         if (layout == NULL)
1410                 return NULL;
1411         else
1412                 return &FILELAYOUT_FROM_HDR(layout)->commit_info;
1413 }
1414
1415 static struct pnfs_layoutdriver_type filelayout_type = {
1416         .id                     = LAYOUT_NFSV4_1_FILES,
1417         .name                   = "LAYOUT_NFSV4_1_FILES",
1418         .owner                  = THIS_MODULE,
1419         .alloc_layout_hdr       = filelayout_alloc_layout_hdr,
1420         .free_layout_hdr        = filelayout_free_layout_hdr,
1421         .alloc_lseg             = filelayout_alloc_lseg,
1422         .free_lseg              = filelayout_free_lseg,
1423         .pg_read_ops            = &filelayout_pg_read_ops,
1424         .pg_write_ops           = &filelayout_pg_write_ops,
1425         .get_ds_info            = &filelayout_get_ds_info,
1426         .mark_request_commit    = filelayout_mark_request_commit,
1427         .clear_request_commit   = filelayout_clear_request_commit,
1428         .scan_commit_lists      = filelayout_scan_commit_lists,
1429         .recover_commit_reqs    = filelayout_recover_commit_reqs,
1430         .search_commit_reqs     = filelayout_search_commit_reqs,
1431         .commit_pagelist        = filelayout_commit_pagelist,
1432         .read_pagelist          = filelayout_read_pagelist,
1433         .write_pagelist         = filelayout_write_pagelist,
1434         .alloc_deviceid_node    = filelayout_alloc_deviceid_node,
1435         .free_deviceid_node     = filelayout_free_deveiceid_node,
1436 };
1437
1438 static int __init nfs4filelayout_init(void)
1439 {
1440         printk(KERN_INFO "%s: NFSv4 File Layout Driver Registering...\n",
1441                __func__);
1442         return pnfs_register_layoutdriver(&filelayout_type);
1443 }
1444
1445 static void __exit nfs4filelayout_exit(void)
1446 {
1447         printk(KERN_INFO "%s: NFSv4 File Layout Driver Unregistering...\n",
1448                __func__);
1449         pnfs_unregister_layoutdriver(&filelayout_type);
1450 }
1451
1452 MODULE_ALIAS("nfs-layouttype4-1");
1453
1454 module_init(nfs4filelayout_init);
1455 module_exit(nfs4filelayout_exit);